diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
new file mode 100644
index 0000000000..04b082bcf9
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
@@ -0,0 +1,132 @@
+name: "kimi-k3-vllm-agg-gb200-dep16-throughput-agentic"
+
+# Day-0 GB200 translation of the official throughput-oriented multi_node_dep
+# profile. TP4 x DP4 gives EP16 across four four-GPU GB200 nodes.
+# https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=4&strategy=multi_node_dep
+
+model:
+  path: "kimi-k3"
+  container: "nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.4.0-kimi-k3-dev.1"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "moonshotai/Kimi-K3"
+  container:
+    image: "nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.4.0-kimi-k3-dev.1"
+  frameworks:
+    dynamo: "1.3.0"
+
+dynamo:
+  install: false
+
+environment:
+  ETCD_LEASE_TTL: "7200"
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  max_attempts: 2160
+  interval_seconds: 10
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  agg_nodes: 4
+  agg_workers: 1
+  gpus_per_agg: 16
+
+infra:
+  etcd_nats_dedicated_node: false
+  nats_max_payload_mb: 32
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  args:
+    dyn-chat-processor: "dynamo"
+    router-mode: "kv"
+    router-kv-events: true
+    router-temperature: "0"
+    router-min-initial-workers: 1
+    kv-cache-block-size: 64
+
+backend:
+  type: vllm
+  connector: null
+  aggregated_environment:
+    HF_HUB_CACHE: "/hf_hub_cache"
+    HUGGINGFACE_HUB_CACHE: "/hf_hub_cache"
+    TRANSFORMERS_CACHE: "/hf_hub_cache"
+    VLLM_ENGINE_READY_TIMEOUT_S: "7200"
+    VLLM_RPC_TIMEOUT: "600000"
+    VLLM_LOG_STATS_INTERVAL: "1"
+    VLLM_USE_V2_MODEL_RUNNER: "0"
+    VLLM_ENABLE_K3_LATENT_MOE_TAIL_FUSION: "1"
+    VLLM_ALLREDUCE_USE_FLASHINFER: "1"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_NET_DEVICES: "mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1"
+    UCX_TLS: "rc,cuda_copy"
+    NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+    NCCL_P2P_LEVEL: "NVL"
+    NVIDIA_GDRCOPY: "1"
+    PYTORCH_ALLOC_CONF: "expandable_segments:True"
+    DG_JIT_CACHE_DIR: "/tmp/dg-cache-kimi-k3-gb200-dep16-{job_id}"
+  vllm_config:
+    aggregated:
+      served-model-name: "moonshotai/Kimi-K3"
+      tensor-parallel-size: 4
+      pipeline-parallel-size: 1
+      data-parallel-size: 4
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      load-format: "fastsafetensors"
+      safetensors-load-strategy: "lazy"
+      kv-cache-dtype: "fp8"
+      attention-backend: "FLASHINFER_MLA"
+      attention-config: '{"mla_prefill_backend":"TRTLLM_RAGGED","use_prefill_query_quantization":true}'
+      moe-backend: "deep_gemm_mega_moe"
+      kda-prefill-backend: "flashkda"
+      kernel-config: '{"enable_cutedsl_warmup":true}'
+      all2all-backend: "flashinfer_nvlink_one_sided"
+      gpu-memory-utilization: 0.92
+      max-num-seqs: 128
+      max-num-batched-tokens: 16384
+      block-size: 64
+      language-model-only: true
+      disable-custom-all-reduce: true
+      enable-prefix-caching: true
+      scheduler-cls: "vllm.v1.core.sched.async_scheduler.AsyncScheduler"
+      dyn-tool-call-parser: "kimi_k3"
+      reasoning-parser: "kimi_k3"
+      dyn-reasoning-parser: "kimi_k3"
+      no-enable-flashinfer-autotune: true
+
+sbatch_directives:
+  cpus-per-task: "144"
+  mem: "0"
+
+srun_options:
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  aiperf_server_metrics: true
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400"
+    AGENTIC_WARMUP_GRACE_PERIOD: "3600"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
new file mode 100644
index 0000000000..6db5821a1a
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
@@ -0,0 +1,131 @@
+name: "kimi-k3-vllm-agg-gb200-tep16-balanced-agentic"
+
+# Day-0 GB200 translation of the official balanced multi_node_tep profile.
+# Dense layers and MoE experts are sharded across 16 GPUs on four GB200 nodes
+# with the official FP8 KV cache.
+# https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=4&strategy=multi_node_tep
+
+model:
+  path: "kimi-k3"
+  container: "nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.4.0-kimi-k3-dev.1"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "moonshotai/Kimi-K3"
+  container:
+    image: "nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.4.0-kimi-k3-dev.1"
+  frameworks:
+    dynamo: "1.3.0"
+
+dynamo:
+  install: false
+
+environment:
+  ETCD_LEASE_TTL: "7200"
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  max_attempts: 2160
+  interval_seconds: 10
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  agg_nodes: 4
+  agg_workers: 1
+  gpus_per_agg: 16
+
+infra:
+  etcd_nats_dedicated_node: false
+  nats_max_payload_mb: 32
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  args:
+    dyn-chat-processor: "dynamo"
+    router-mode: "kv"
+    router-kv-events: true
+    router-temperature: "0"
+    router-min-initial-workers: 1
+    kv-cache-block-size: 64
+
+backend:
+  type: vllm
+  connector: null
+  aggregated_environment:
+    HF_HUB_CACHE: "/hf_hub_cache"
+    HUGGINGFACE_HUB_CACHE: "/hf_hub_cache"
+    TRANSFORMERS_CACHE: "/hf_hub_cache"
+    VLLM_ENGINE_READY_TIMEOUT_S: "7200"
+    VLLM_RPC_TIMEOUT: "600000"
+    VLLM_LOG_STATS_INTERVAL: "1"
+    VLLM_USE_V2_MODEL_RUNNER: "0"
+    VLLM_ENABLE_K3_LATENT_MOE_TAIL_FUSION: "1"
+    VLLM_ALLREDUCE_USE_FLASHINFER: "1"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_NET_DEVICES: "mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1"
+    UCX_TLS: "rc,cuda_copy"
+    NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+    NCCL_P2P_LEVEL: "NVL"
+    NVIDIA_GDRCOPY: "1"
+    PYTORCH_ALLOC_CONF: "expandable_segments:True"
+    DG_JIT_CACHE_DIR: "/tmp/dg-cache-kimi-k3-gb200-tep16-{job_id}"
+  vllm_config:
+    aggregated:
+      served-model-name: "moonshotai/Kimi-K3"
+      tensor-parallel-size: 16
+      pipeline-parallel-size: 1
+      enable-expert-parallel: true
+      trust-remote-code: true
+      load-format: "fastsafetensors"
+      safetensors-load-strategy: "lazy"
+      kv-cache-dtype: "fp8"
+      attention-backend: "FLASHINFER_MLA"
+      attention-config: '{"mla_prefill_backend":"TRTLLM_RAGGED","use_prefill_query_quantization":true}'
+      moe-backend: "flashinfer_trtllm"
+      kda-prefill-backend: "flashkda"
+      kernel-config: '{"enable_cutedsl_warmup":true}'
+      all2all-backend: "flashinfer_nvlink_one_sided"
+      gpu-memory-utilization: 0.92
+      max-num-seqs: 32
+      max-num-batched-tokens: 8192
+      block-size: 64
+      language-model-only: true
+      disable-custom-all-reduce: true
+      enable-prefix-caching: true
+      scheduler-cls: "vllm.v1.core.sched.async_scheduler.AsyncScheduler"
+      dyn-tool-call-parser: "kimi_k3"
+      reasoning-parser: "kimi_k3"
+      dyn-reasoning-parser: "kimi_k3"
+      no-enable-flashinfer-autotune: true
+      compilation-config: '{"pass_config":{"fuse_allreduce_rms":false}}'
+
+sbatch_directives:
+  cpus-per-task: "144"
+  mem: "0"
+
+srun_options:
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  aiperf_server_metrics: true
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
new file mode 100644
index 0000000000..8c5389505b
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
@@ -0,0 +1,128 @@
+name: "kimi-k3-vllm-agg-gb200-tp16-latency-agentic"
+
+# Day-0 GB200 translation of the official latency-oriented multi_node_tp
+# profile. TP16 spans four GB200 nodes and uses the official FP8 KV cache.
+# https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=4&strategy=multi_node_tp
+
+model:
+  path: "kimi-k3"
+  container: "nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.4.0-kimi-k3-dev.1"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "moonshotai/Kimi-K3"
+  container:
+    image: "nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.4.0-kimi-k3-dev.1"
+  frameworks:
+    dynamo: "1.3.0"
+
+dynamo:
+  install: false
+
+environment:
+  ETCD_LEASE_TTL: "7200"
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  max_attempts: 2160
+  interval_seconds: 10
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  agg_nodes: 4
+  agg_workers: 1
+  gpus_per_agg: 16
+
+infra:
+  etcd_nats_dedicated_node: false
+  nats_max_payload_mb: 32
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  args:
+    dyn-chat-processor: "dynamo"
+    router-mode: "kv"
+    router-kv-events: true
+    router-temperature: "0"
+    router-min-initial-workers: 1
+    kv-cache-block-size: 64
+
+backend:
+  type: vllm
+  connector: null
+  aggregated_environment:
+    HF_HUB_CACHE: "/hf_hub_cache"
+    HUGGINGFACE_HUB_CACHE: "/hf_hub_cache"
+    TRANSFORMERS_CACHE: "/hf_hub_cache"
+    VLLM_ENGINE_READY_TIMEOUT_S: "7200"
+    VLLM_RPC_TIMEOUT: "600000"
+    VLLM_LOG_STATS_INTERVAL: "1"
+    VLLM_USE_V2_MODEL_RUNNER: "0"
+    VLLM_ENABLE_K3_LATENT_MOE_TAIL_FUSION: "1"
+    VLLM_ALLREDUCE_USE_FLASHINFER: "1"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_NET_DEVICES: "mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1"
+    UCX_TLS: "rc,cuda_copy"
+    NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+    NCCL_P2P_LEVEL: "NVL"
+    NVIDIA_GDRCOPY: "1"
+    PYTORCH_ALLOC_CONF: "expandable_segments:True"
+    DG_JIT_CACHE_DIR: "/tmp/dg-cache-kimi-k3-gb200-tp16-{job_id}"
+  vllm_config:
+    aggregated:
+      served-model-name: "moonshotai/Kimi-K3"
+      tensor-parallel-size: 16
+      pipeline-parallel-size: 1
+      trust-remote-code: true
+      load-format: "fastsafetensors"
+      safetensors-load-strategy: "lazy"
+      kv-cache-dtype: "fp8"
+      attention-backend: "FLASHINFER_MLA"
+      attention-config: '{"mla_prefill_backend":"TRTLLM_RAGGED","use_prefill_query_quantization":true}'
+      moe-backend: "flashinfer_trtllm"
+      kda-prefill-backend: "flashkda"
+      kernel-config: '{"enable_cutedsl_warmup":true}'
+      gpu-memory-utilization: 0.92
+      max-num-seqs: 8
+      max-num-batched-tokens: 8192
+      block-size: 64
+      language-model-only: true
+      disable-custom-all-reduce: true
+      enable-prefix-caching: true
+      scheduler-cls: "vllm.v1.core.sched.async_scheduler.AsyncScheduler"
+      dyn-tool-call-parser: "kimi_k3"
+      reasoning-parser: "kimi_k3"
+      dyn-reasoning-parser: "kimi_k3"
+      no-enable-flashinfer-autotune: true
+      compilation-config: '{"pass_config":{"fuse_allreduce_rms":false}}'
+
+sbatch_directives:
+  cpus-per-task: "144"
+  mem: "0"
+
+srun_options:
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  aiperf_server_metrics: true
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 8cc99a8a0b..9ad4926354 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8049,6 +8049,78 @@ dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg:
           tp: 8
           ep: 8
           dp-attn: true
+
+# Day-0 Kimi K3 AgentX coverage on GB200. The three aggregate srt-slurm
+# profiles mirror the serving strategies synthesized by the official recipe:
+#   latency:    multi-node TP16       (4 GB200 nodes / 16 GPUs)
+#   balanced:   multi-node TEP16      (4 GB200 nodes / 16 GPUs)
+#   throughput: multi-node TP4 x DP4  (4 GB200 nodes / EP16)
+# https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200
+kimik3-fp4-gb200-dynamo-vllm-agentic:
+  image: nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.4.0-kimi-k3-dev.1
+  model: moonshotai/Kimi-K3
+  model-prefix: kimik3
+  runner: cluster:gb200-nv
+  precision: fp4
+  framework: dynamo-vllm
+  router: { name: dynamo-router, version: "1.3.0" }
+  multinode: true
+  disagg: false
+  scenarios:
+    agentic-coding:
+    - search-space:
+      # Latency oriented: multi_node_tp strategy, TP16 across four GB200
+      # nodes.
+      # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=4&strategy=multi_node_tp
+      - spec-decoding: none
+        conc-list: [1, 2, 4, 8]
+        prefill:
+          num-worker: 1
+          tp: 16
+          ep: 16
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml"
+        decode:
+          num-worker: 0
+          tp: 16
+          ep: 16
+          dp-attn: false
+      # Balanced: multi_node_tep strategy, TEP16 across four GB200 nodes.
+      # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=4&strategy=multi_node_tep
+      - spec-decoding: none
+        conc-list: [8, 16, 24, 32]
+        prefill:
+          num-worker: 1
+          tp: 16
+          ep: 16
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml"
+        decode:
+          num-worker: 0
+          tp: 16
+          ep: 16
+          dp-attn: false
+      # Throughput oriented: official multi_node_dep strategy, DEP16 across
+      # four GB200 nodes (TP4 x DP4 = EP16, one local DP rank per node).
+      # The recipe allows up to 3600s for full-context saturation warmup to drain.
+      # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=4&strategy=multi_node_dep
+      - spec-decoding: none
+        conc-list: [32, 64, 96, 128, 192, 256]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 16
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml"
+        decode:
+          num-worker: 0
+          tp: 4
+          ep: 16
+          dp-attn: true
+
 dsv4-fp4-gb300-dynamo-sglang-agentic-agg:
   image: lmsysorg/sglang:nightly-dev-cu13-20260711-7de33ce8
   model: deepseek-ai/DeepSeek-V4-Pro
diff --git a/configs/runners.yaml b/configs/runners.yaml
index 851b821ba2..bb0156ab9a 100644
--- a/configs/runners.yaml
+++ b/configs/runners.yaml
@@ -247,6 +247,7 @@ labels:
     - gb200-nv_0
     - gb200-nv_1
     - gb200-nv_2
+    - gb200-nv_3
   cluster:gb300-nv:
     - gb300-nv_0
     - gb300-nv_1
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 09108c33ea..7315dd0060 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5159,3 +5159,10 @@
     - "Bring-up validated in run 30326393603: all 12 configs green, zero ServerDisconnectedError after the keep-alive fix. GPU KV resolves to 42.23 GiB / 3,249,215 tokens, i.e. ~3.1 max-length requests, against MAX_NUM_SEQS = 2*CONC."
     - "Measured behaviour: below conc 8 the GPU-resident and DRAM arms are within run-to-run noise (1-5%). At conc 16 and 24 the GPU-resident arm thrashes -- prefix cache hit rate 2.7%, TTFT p50 86s and 191s, 49.6 and 54.9 output tok/s -- because the working set exceeds GPU KV and prefixes are recomputed. The DRAM arm holds TTFT p50 0.85s and 6.2s for 245.0 and 260.6 output tok/s (4-5x), with the CPU tier serving a 62% external prefix cache hit rate at conc 24. The high-conc GPU-resident points are retained deliberately as the honest baseline that makes the offload gain legible."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2371
+
+
+- config-keys:
+    - kimik3-fp4-gb200-dynamo-vllm-agentic
+  description:
+    - "Add Day-0 Kimi K3 MXFP4 GB200 AgentX performance points following the official vLLM recipe: TP16 (conc 1-8), TEP16 (conc 8-32), and TP4 x DP4 = EP16 (conc 32-256). https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2354
diff --git a/runners/launch_gb200-nv.sh b/runners/launch_gb200-nv.sh
index 426cd2dd87..2f913e692c 100755
--- a/runners/launch_gb200-nv.sh
+++ b/runners/launch_gb200-nv.sh
@@ -155,6 +155,12 @@ elif [[ $FRAMEWORK == "dynamo-vllm" ]]; then
     if [[ $MODEL_PREFIX == "kimik2.5" && $PRECISION == "fp4" ]]; then
         export MODEL_PATH="/mnt/lustre01/models/kimi-k2.5-nvfp4"
         export SRT_SLURM_MODEL_PREFIX="kimi-k2.5-nvfp4"
+    elif [[ $MODEL_PREFIX == "kimik3" && $PRECISION == "fp4" ]]; then
+        # Load Kimi K3 from node-local NVMe for faster startup. The checkpoint
+        # must be pre-staged at this exact path on every allocated GB200 node.
+        # This alias matches model.path in the checked-in AgentX recipes.
+        export MODEL_PATH="/mnt/numa1/models/Kimi-K3"
+        export SRT_SLURM_MODEL_PREFIX="kimi-k3"
     elif [[ $MODEL_PREFIX == "dsv4" && $PRECISION == "fp4" ]]; then
         # The FP4 checkpoint is staged on compute-visible Lustre. The former
         # /mnt/numa1 path is no longer present on watchtower compute nodes;
@@ -172,7 +178,7 @@ elif [[ $FRAMEWORK == "dynamo-vllm" ]]; then
         export MODEL_PATH="/mnt/lustre01/models/MiniMax-M3-MXFP8"
         export SRT_SLURM_MODEL_PREFIX="minimax-m3-mxfp8"
     else
-        echo "Unsupported model prefix/precision combination: $MODEL_PREFIX/$PRECISION. Supported combinations for dynamo-vllm: kimik2.5/fp4, dsv4/fp4, minimaxm2.5/fp4, minimaxm2.5/fp8, minimaxm3/fp8"
+        echo "Unsupported model prefix/precision combination: $MODEL_PREFIX/$PRECISION. Supported combinations for dynamo-vllm: kimik2.5/fp4, kimik3/fp4, dsv4/fp4, minimaxm2.5/fp4, minimaxm2.5/fp8, minimaxm3/fp8"
         exit 1
     fi
 else
@@ -183,7 +189,7 @@ NGINX_IMAGE="nginx:1.27.4"
 
 uses_watchtower_shared_fs() {
     case "$MODEL_PREFIX" in
-        minimaxm2.5|minimaxm3|kimik2.5|qwen3.5) return 0 ;;
+        minimaxm2.5|minimaxm3|kimik2.5|kimik3|qwen3.5) return 0 ;;
         *) return 1 ;;
     esac
 }
@@ -352,8 +358,7 @@ fi
 
 # TODO(CJQ): make first class upon srt-slurm upstream refactor
 if [[ "$IS_AGENTIC" == "1" ]]; then
-    # Agentic multi-node uses the same pinned cquil11/srt-slurm-nv commit as
-    # launch_gb300-nv.sh — everything the agentic recipes need is there:
+    # Agentic multi-node pins cquil11/srt-slurm-nv revisions that provide:
     #   - BenchmarkType.CUSTOM + benchmark.command + benchmark.env
     #     (the hook that hands off to benchmarks/multi_node/agentic_srt.sh)
     #   - DynamoConfig.wheel (recipes pin the ai-dynamo wheel)
@@ -363,10 +368,19 @@ if [[ "$IS_AGENTIC" == "1" ]]; then
     #     must reach the agentic_srt.sh srun)
     git clone https://github.com/cquil11/srt-slurm-nv.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR"
-    git checkout de59739b172e507e15ebf145bfe305f606e82fbf
-    mkdir -p recipes/vllm/deepseek-v4/agentic
-    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic" \
-        recipes/vllm/deepseek-v4/agentic
+    if [[ "$MODEL_PREFIX" == "kimik3" ]]; then
+        # Kimi K3 additionally needs vLLM TP groups within data parallel and
+        # every DP engine metrics endpoint exposed to AIPerf.
+        git checkout b1fb626fbdbfe3306dcb51cb181ab35861ec3b1c
+        mkdir -p recipes/vllm/kimi-k3/agentic
+        cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic" \
+            recipes/vllm/kimi-k3/agentic
+    else
+        git checkout de59739b172e507e15ebf145bfe305f606e82fbf
+        mkdir -p recipes/vllm/deepseek-v4/agentic
+        cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic" \
+            recipes/vllm/deepseek-v4/agentic
+    fi
 elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "dsv4" ]]; then
     git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR"
@@ -630,43 +644,54 @@ if [[ "${EVAL_ONLY:-false}" != "true" ]]; then
         exit 1
     fi
 
-    # Find all result subdirectories
-    RESULT_SUBDIRS=$(find "$LOGS_DIR" -maxdepth 1 -type d -name "*isl*osl*" 2>/dev/null)
-
-    if [ -z "$RESULT_SUBDIRS" ]; then
-        echo "Warning: No result subdirectories found in $LOGS_DIR"
+    if [[ "$IS_AGENTIC" == "1" ]]; then
+        # The custom benchmark runs inside the compute-visible
+        # INFMAX_WORKSPACE mount. Its aggregation step writes one
+        # ${RESULT_FILENAME}_conc<N>.json there per point; stage those files
+        # back to GITHUB_WORKSPACE for the workflow guard and artifact upload.
+        copy_agentic_results \
+            "$INFMAX_WORKSPACE" \
+            "$GITHUB_WORKSPACE" \
+            "$RESULT_FILENAME" || exit 1
     else
-        # Process results from all configurations
-        for result_subdir in $RESULT_SUBDIRS; do
-            echo "Processing result subdirectory: $result_subdir"
-
-            # Extract configuration info from directory name
-            CONFIG_NAME=$(basename "$result_subdir")
-
-            # Find all result JSON files
-            RESULT_FILES=$(find "$result_subdir" -name "results_concurrency_*.json" 2>/dev/null)
-
-            for result_file in $RESULT_FILES; do
-                if [ -f "$result_file" ]; then
-                    # Extract metadata from filename
-                    # Files may be "results_concurrency_N_gpus_G_ctx_C_gen_D.json" (disagg) or "results_concurrency_N_gpus_G.json" (non-disagg)
-                    filename=$(basename "$result_file")
-                    concurrency=$(echo "$filename" | sed -n 's/results_concurrency_\([0-9]*\)_gpus_.*/\1/p')
-                    gpus=$(echo "$filename" | sed -n 's/results_concurrency_[0-9]*_gpus_\([0-9][0-9]*\).*/\1/p')
-                    ctx=$(echo "$filename" | sed -n 's/.*_ctx_\([0-9]*\)_gen_.*/\1/p')
-                    gen=$(echo "$filename" | sed -n 's/.*_gen_\([0-9]*\)\.json/\1/p')
-
-                    echo "Processing concurrency $concurrency with $gpus GPUs (ctx: $ctx, gen: $gen): $result_file"
-
-                    if [ -n "$ctx" ] && [ -n "$gen" ]; then
-                        WORKSPACE_RESULT_FILE="$GITHUB_WORKSPACE/${RESULT_FILENAME}_${CONFIG_NAME}_conc${concurrency}_gpus_${gpus}_ctx_${ctx}_gen_${gen}.json"
-                    else
-                        WORKSPACE_RESULT_FILE="$GITHUB_WORKSPACE/${RESULT_FILENAME}_${CONFIG_NAME}_conc${concurrency}_gpus_${gpus}.json"
+        # Find all fixed-sequence result subdirectories.
+        RESULT_SUBDIRS=$(find "$LOGS_DIR" -maxdepth 1 -type d -name "*isl*osl*" 2>/dev/null)
+
+        if [ -z "$RESULT_SUBDIRS" ]; then
+            echo "Warning: No result subdirectories found in $LOGS_DIR"
+        else
+            # Process results from all configurations
+            for result_subdir in $RESULT_SUBDIRS; do
+                echo "Processing result subdirectory: $result_subdir"
+
+                # Extract configuration info from directory name
+                CONFIG_NAME=$(basename "$result_subdir")
+
+                # Find all result JSON files
+                RESULT_FILES=$(find "$result_subdir" -name "results_concurrency_*.json" 2>/dev/null)
+
+                for result_file in $RESULT_FILES; do
+                    if [ -f "$result_file" ]; then
+                        # Extract metadata from filename
+                        # Files may be "results_concurrency_N_gpus_G_ctx_C_gen_D.json" (disagg) or "results_concurrency_N_gpus_G.json" (non-disagg)
+                        filename=$(basename "$result_file")
+                        concurrency=$(echo "$filename" | sed -n 's/results_concurrency_\([0-9]*\)_gpus_.*/\1/p')
+                        gpus=$(echo "$filename" | sed -n 's/results_concurrency_[0-9]*_gpus_\([0-9][0-9]*\).*/\1/p')
+                        ctx=$(echo "$filename" | sed -n 's/.*_ctx_\([0-9]*\)_gen_.*/\1/p')
+                        gen=$(echo "$filename" | sed -n 's/.*_gen_\([0-9]*\)\.json/\1/p')
+
+                        echo "Processing concurrency $concurrency with $gpus GPUs (ctx: $ctx, gen: $gen): $result_file"
+
+                        if [ -n "$ctx" ] && [ -n "$gen" ]; then
+                            WORKSPACE_RESULT_FILE="$GITHUB_WORKSPACE/${RESULT_FILENAME}_${CONFIG_NAME}_conc${concurrency}_gpus_${gpus}_ctx_${ctx}_gen_${gen}.json"
+                        else
+                            WORKSPACE_RESULT_FILE="$GITHUB_WORKSPACE/${RESULT_FILENAME}_${CONFIG_NAME}_conc${concurrency}_gpus_${gpus}.json"
+                        fi
+                        copy_to_workspace "$result_file" "$WORKSPACE_RESULT_FILE" || exit 1
                     fi
-                    copy_to_workspace "$result_file" "$WORKSPACE_RESULT_FILE" || exit 1
-                fi
+                done
             done
-        done
+        fi
     fi
 
     echo "All result files processed"
diff --git a/runners/slurm_utils.sh b/runners/slurm_utils.sh
index c26b4f1690..55c8652b84 100644
--- a/runners/slurm_utils.sh
+++ b/runners/slurm_utils.sh
@@ -41,6 +41,36 @@ copy_to_workspace() {
     echo "Copied $(basename "$source_file") to $destination_file"
 }
 
+copy_agentic_results() {
+    local source_dir="$1"
+    local workspace="$2"
+    local result_filename="$3"
+    local result_file
+    local copied=0
+
+    if [[ ! -d "$source_dir" ]]; then
+        echo "ERROR: agentic result directory not found at $source_dir" >&2
+        return 1
+    fi
+
+    while IFS= read -r -d '' result_file; do
+        copy_to_workspace \
+            "$result_file" \
+            "$workspace/$(basename "$result_file")" || return 1
+        copied=$((copied + 1))
+    done < <(
+        find "$source_dir" -maxdepth 1 -type f \
+            -name "${result_filename}_conc*.json" -print0
+    )
+
+    if [[ "$copied" -eq 0 ]]; then
+        echo "ERROR: no ${result_filename}_conc*.json results found in $source_dir" >&2
+        return 1
+    fi
+
+    echo "Copied $copied agentic result file(s)"
+}
+
 copy_eval_artifacts() {
     local eval_dir="$1"
     local workspace="$2"
diff --git a/runners/test_slurm_utils.py b/runners/test_slurm_utils.py
new file mode 100644
index 0000000000..c027cc9b2f
--- /dev/null
+++ b/runners/test_slurm_utils.py
@@ -0,0 +1,57 @@
+import subprocess
+from pathlib import Path
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+SLURM_UTILS = REPO_ROOT / "runners" / "slurm_utils.sh"
+
+
+def run_bash(command: str, *args: Path | str) -> subprocess.CompletedProcess[str]:
+    return subprocess.run(
+        ["bash", "-c", command, "bash", *(str(arg) for arg in args)],
+        check=False,
+        capture_output=True,
+        text=True,
+    )
+
+
+def test_copy_agentic_results_stages_only_matching_points(tmp_path: Path) -> None:
+    source = tmp_path / "source"
+    workspace = tmp_path / "workspace"
+    source.mkdir()
+    workspace.mkdir()
+    (source / "run_conc1.json").write_text('{"conc": 1}\n')
+    (source / "run_conc16.json").write_text('{"conc": 16}\n')
+    (source / "other_conc1.json").write_text('{"conc": 1}\n')
+
+    result = run_bash(
+        'source "$1"; copy_agentic_results "$2" "$3" run',
+        SLURM_UTILS,
+        source,
+        workspace,
+    )
+
+    assert result.returncode == 0, result.stderr
+    assert sorted(path.name for path in workspace.iterdir()) == [
+        "run_conc1.json",
+        "run_conc16.json",
+    ]
+
+
+def test_copy_agentic_results_fails_when_aggregate_is_missing(
+    tmp_path: Path,
+) -> None:
+    source = tmp_path / "source"
+    workspace = tmp_path / "workspace"
+    source.mkdir()
+    workspace.mkdir()
+
+    result = run_bash(
+        'source "$1"; copy_agentic_results "$2" "$3" run',
+        SLURM_UTILS,
+        source,
+        workspace,
+    )
+
+    assert result.returncode != 0
+    assert "no run_conc*.json results found" in result.stderr